Same prompt
Every model in a test gets the exact same prompt. The input never moves — that's the control.
A YouTube series · on-device AI benchmarks
PC Magik Lab
> Local models, real GPUs.
A hands-on series about running AI models on real, local GPUs. One machine, one method, every prompt and every result file in the open — so you can see exactly what local AI can and cannot do.
Methodology
01
02
03
04
05
06
How the series measures things
No cherry-picked runs, no mystery settings. Every test in the series follows the same controlled process so the only thing that changes is the thing being tested.
Same rig
One single machine for the whole series. Same GPU, same server, same conditions from run to run.
One variable per test
Each test changes exactly one variable — so any difference in the output is meaningful, not noise.
Model reloaded before every run
The model is reloaded fresh before every run, so no leftover state ever skews a result.
Everything published
Every prompt and every result file goes straight to the public repository. Reproduce it yourself.
Local, not the cloud
Everything runs on local hardware. No remote inference — the whole point is what your own GPU can do.
The full set of prompts and result files for every test is published in the repository — open it and verify any claim the series makes.
The rig
One rig. No excuses.
The entire series runs on a single workstation. One GPU, one model server — a fixed environment where the only thing in motion is the variable under test.
- gpuNVIDIA RTX 3090
- vram24 GB
- rigs1 — always this one
Model server: LM Studio
Models are served locally with LM Studio, the local model server the whole series is built around. Same server, same machine, every single test.
fixed prompt
model reload
LM Studio
RTX 3090
result file → repo
The pipeline is deliberately boring: load, run, publish. When the setup this simple still produces different answers between models, that's a signal worth showing on camera.
About
Who runs the lab
PC Magik Lab is the work of PC Magik (Mateusz Piekut) — a computer service and homelab operator from Poland who spends as much time inside machines as in front of them. The series turns that daily, hands-on relationship with real hardware into something you can watch: local AI models, run on real GPUs, measured without drama.
YouTube channel
Watch the series — every test, on camera, on real hardware.
youtube.com/@PCMagikLabPrompts & result files
The public repository with every prompt and every result file, ready to inspect.
github.com/pcmagik/pcmagik-labFrom Poland
Recorded and run by a computer service and homelab operator based in Poland.